A study enhances physics-constrained neural networks by introducing an upgraded numerical solver, a unified autoregressive block, and two neural backbones. These improvements reduce root mean squared error by 8-22% in short-term forecasts over the South Pacific and better preserve physical consistency.